TuskPointGitHub

Documentation

What is TuskPoint?

TuskPoint is a drop-in LangGraph checkpointer that stores every step of an agent run as an immutable Walrus blob. When a process crashes, you resume from exactly where it stopped — not from the beginning — and you can fork, diff, search, and audit the entire history.

The problem

Long-running and multi-agent workflows accumulate state: research notes, intermediate artifacts, tool outputs, decisions. When the process dies — a timeout, an OOM, a redeploy — that state usually dies with it, and the agent restarts from zero. Keeping it in a local file or a single database row makes it fragile and hard to inspect.

TuskPoint persists each checkpoint to Walrus, a decentralized blob store where every blob is content-addressed. A read returns the exact bytes you wrote, verifiably, from any aggregator — no central server to trust and nothing to lose on a crash.

How the pieces fit

Three layers, one drop-in saver. LangGraph calls the standard checkpoint API; TuskPoint serializes and gzips the state, stores it on Walrus, and records the blob in a per-thread manifest. MemWal adds a one-line natural-language summary so you can recall checkpoints in plain English. An MCP server exposes the whole thing as eight tools any agent can call.

  • WalrusSaver — the exact, byte-for-byte layer. The source of truth you rewind to.
  • MemWal — the semantic index. Pointers into the exact store, never the truth itself.
  • MCP server — the all-in-one toolbelt: checkpoint_save, checkpoint_fork, verify_trail, and more.

Exact vs. semantic — both, on purpose

Lookups by ID are deterministic and content-addressed. Semantic search is only for discovery: it returns summaries carrying checkpoint IDs you then load exactly. Vector recall indexes the exact store; it is never the source of truth.

Keep reading